From sugar mills and power plants to AI data centers, expensive infrastructure gets built by selling its future output before it exists.

Bankability is just the art of borrowing against demand that hasn’t happened yet.

The more I read about the AI infrastructure boom, the more it looks like one vast mutual-hostage situation: a developer who can’t build without capital, a bank that won’t lend without a buyer, and a buyer who won’t commit to an asset that doesn’t exist yet. The only way anyone moves first is if everyone commits on paper.

Paper Before Steam

Two hundred years ago, a version of that deadlock hit France. When blockades during the Napoleonic Wars severed France from West Indian cane sugar, French entrepreneurs rushed to industrialize sugar extraction from domestic beets. The process was becoming technically possible, but the economics were brutal since a sugar mill required heavy steam boilers, crushing machinery, and massive debt. Beets, however, rot quickly after harvest and are too heavy to transport long distances.

If a factory was built without a local crop, the machinery starved and the owner went bankrupt. If local farmers planted beets without a factory nearby, their harvest rotted in the field. Both sides were mutual hostages to an asset that didn’t exist yet.

What broke the deadlock was paper: multi-year crop delivery contracts. Factory operators signed forward off-take agreements with surrounding landowners, guaranteeing a floor price per ton of beets. Factory owners could then take those signed farmer commitments to lenders as proof of guaranteed feedstock to secure debt for the steam engines.

From Gas Wells to Power Plants

I didn’t fully appreciate how deeply this paper loop governs the modern world until I got into bitcoin mining and began building energy-intensive infrastructure myself.  Back then, I didn’t really understand the business of energy. It felt a little Byzantine, with subtle nods to financiers, make-believe forecasting of demand, and an entirely new dictionary for describing fairly simple things.  

Take Power Purchase Agreements, or PPAs. The name is straightforward enough, but the terminology used is obscure and only makes sense with historical context.

Take-or-pay: the buyer commits to paying for a minimum volume of power over a long-term contract (often 15 to 20 years), whether they use it or not.

Take-and-pay: the buyer only pays for the power they actually consume, so the risk of unused generation stays with the producer.

The “take-or-pay” structure originated in the simple language of the 1930s American natural gas country. A gas well in Texas had exactly one potential buyer: the physical pipeline connected to it. Because gas couldn’t be stored at the wellhead, a producer had to dedicate their entire output to that single pipe. In return, the pipeline promised to take a minimum volume every year or pay for it anyway. By the 1950s, this risk split was boilerplate finance. 

In emerging markets, no grid-scale renewable project gets financed without first locking in a long-term PPA with the national utility. And the industry standard remains take-or-pay whether the grid needs it or not.

This is a heavy commitment on both sides. Developers spend years and a lot of money chasing a PPA before they can break ground, while utilities sign up for decades of payments on power they may never use. So projects move slower, the grid loses flexibility, and everyone ends up paying more for electricity.  In short: banks won’t finance an expensive capital build without a readymade, creditworthy buyer locked in for years.

Turns out this is how it works with AI data centers too. A PPA makes a power plant bankable by selling tomorrow’s electricity before the plant is built. AI infrastructure is beginning to use the same trick: selling tomorrow’s compute before the GPUs are installed.

PPAs All the Way Down

When I started digging into how AI data centers were being built, the pattern was hard to miss. Hyperscalers and AI labs were buying “reserved compute,” and sure enough, take-or-pay paper was right at the center of the deals.

The industry calls this a “Compute Offtake Agreement”, or as my friend Joseph Nganga framed it, a Compute Purchase Agreement (CPA). In a CPA, an investment-grade hyperscaler steps into the  same role the national utility plays in a PPA: the creditworthy counterparty that lenders are willing to underwrite.

What makes this model so fascinating is that it isn’t just a single contract… it’s PPAs all the way down, where one company’s off-take agreement becomes the foundation for another’s bankability:

The building and power layer: CoreWeave signs a long-term lease for powered data-center space from an operator like Core Scientific, which brings the building and the grid connection. For Core Scientific, that lease plays the role of a PPA: it pre-sells years of capacity, giving lenders the contracted revenue they need to finance the concrete and megawatts.

The compute layer: Inside the building, CoreWeave repeats the same play downstream. Microsoft reserves the compute that CoreWeave will install, under the kind of long-term contract CoreWeave typically structures on a take-or-pay basis. From CoreWeave’s perspective, this CPA plays the role of a PPA. The terms allocate the demand risk; Microsoft’s signature makes the promise valuable to lenders.

The capital layer: With the building, power and customer lined up, CoreWeave can finance the GPU infrastructure required to deliver the compute. In 2024, it secured commitments for a $7.5 billion debt facility backed by both the physical infrastructure and the contracted cash flows behind it. The GPUs gave lenders collateral and the customer contracts made repayment believable.

Once the capacity is operating and generating revenue, the risk falls and the capital can become cheaper. CoreWeave can refinance expensive build-stage debt and invest again. Build, de-risk, refinance, repeat. Do it enough times and you go public.

The Five-Year Problem

Where the analogy breaks is time. For example a geothermal plant runs for 40 years, so a 20-year PPA still leaves valuable years on the machine. A GPU is old in five. That’s why CPAs run a quarter the length of PPAs, and why the central underwriting question is what the hardware will be worth on the day the contract ends.

The industry’s answer was Nvidia lining up Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build financing platforms targeting $500B+, with Nvidia itself backstopping a slice of the residual value on some deals. Nvidia is putting its balance sheet behind hardware that its own product roadmap will age. It profits when customers upgrade, but on supported deals it may also absorb part of the loss if yesterday’s GPUs lose value too quickly. Nvidia is, in effect, writing insurance on its own upgrade cycle.

In energy, when the offtaker’s credit isn’t enough, a guarantor steps in. Same move here, with the chip vendor as guarantor of last resort. Billions are already flowing into AI compute. Nvidia’s backstop should make that capital cheaper, longer-dated, and available at much greater scale.

Paper Leads to Impossible Futures

What I love is this tension between the courage it takes to make these financial bets and the possibility of both failure and success. Maybe the most compelling realization is that even in failures we still get something positive. 

For instance, during the gas shortages of the 1970s, some US pipelines signed contracts covering as much as 80-90% of a producer’s deliverability. When demand collapsed in the 1980s, pipelines were left holding an estimated $10 billion in take-or-pay liabilities for gas nobody wanted, and it took a decade of lawsuits to unwind. But, even with paper failure, the pipelines themselves kept running, and they’re still moving gas today.

Long-term off-take paper doesn’t make risk disappear, it concentrates it into the credit and demand forecast of the buyer. Bankability simply means someone credible has signed their name beneath the uncertainty.

Yet, when those forecasts hold, this mechanism is what pulls impossible futures into reality. Two hundred years ago, sugar mills and farmers used paper to bind themselves together and build a brand-new agricultural industry from scratch. A century later, the same underlying move of guaranteeing future revenue helped finance early submarine telegraph cables and generations of power plants.  

Now the technology industry is using the same move to build AI capacity by selling tomorrow’s compute before today’s GPUs have been installed.

It’s a high-stakes bet. A take-or-pay contract can’t guarantee the future, and if demand falters, everyone holding the paper feels it. But… if the demand holds, the mutual hostage dynamic does exactly what it was designed to do as it turns an unbankable risk into an extraordinary success with a physical foundation for a new era.

That is how impossible futures arrive: first as a signature, then as concrete and silicon.